A bank's customer chatbot occasionally quotes fee amounts that appear in no bank document. Its prompts are already clear and well structured, and the tone of its answers is fine. Which practice most directly reduces these fabricated figures?
Choose one.
Grounding connects a model's responses to verifiable, authoritative data so that factual claims come from a source rather than from the model's trained memory. It is Google Cloud's main remedy for hallucinated facts.
The problem is invented facts, and the stem rules out a prompting fix. Grounding the chatbot in the official fee schedule gives it the real numbers at answer time. Fine-tuning on old chats changes behaviour, not factual accuracy; few-shot examples change format; drift monitoring only observes.
- Classify the failure: invented figures, which is a hallucination.
- Note the stated constraint: prompts and tone are already fine.
- Pick the practice that supplies verifiable facts at answer time: grounding.
- Discard practices that change style (tuning, examples) or only observe (monitoring).
Exam tip: When a model invents facts that should come from your records, ground it in those records.
Overcoming Foundation Model Limitations: Grounding, RAG, and HITL — the lesson that teaches this.